BTC $63,991.37 -1.34%
ETH $1,790.86 -1.54%
BNB $587.80 -0.84%
XRP $0.9797 +0.98%
SOL $74.11 +0.57%
TRX $0.3096 +1.38%
DOGE $0.0590 -0.59%
ADA $0.1524 -1.40%
BCH $207.22 -1.95%
LINK $9.11 -1.93%
HYPE $68.59 -1.68%
AAVE $110.76 -2.95%
SUI $0.7213 -1.94%
XLM $0.1398 -0.70%
ZEC $929.75 -1.80%
AAPL $350.97 -0.32%
AMZN $259.08 -0.06%
GOOGL $282.83 -0.26%
MSFT $505.67 -2.08%
META $618.33 -0.85%
NVDA $160.01 +1.10%
TSLA $382.60 +0.30%
SNDK $814.68 -1.16%
INTC $55.39 +0.38%
SPCX $142.25 +2.51%
MU $725.52 -0.94%
AMD $511.50 -0.66%
BTC $63,991.37 -1.34%
ETH $1,790.86 -1.54%
BNB $587.80 -0.84%
XRP $0.9797 +0.98%
SOL $74.11 +0.57%
TRX $0.3096 +1.38%
DOGE $0.0590 -0.59%
ADA $0.1524 -1.40%
BCH $207.22 -1.95%
LINK $9.11 -1.93%
HYPE $68.59 -1.68%
AAVE $110.76 -2.95%
SUI $0.7213 -1.94%
XLM $0.1398 -0.70%
ZEC $929.75 -1.80%
AAPL $350.97 -0.32%
AMZN $259.08 -0.06%
GOOGL $282.83 -0.26%
MSFT $505.67 -2.08%
META $618.33 -0.85%
NVDA $160.01 +1.10%
TSLA $382.60 +0.30%
SNDK $814.68 -1.16%
INTC $55.39 +0.38%
SPCX $142.25 +2.51%
MU $725.52 -0.94%
AMD $511.50 -0.66%

h200

All
Article
Flash

first_img Online materials: Ascend 950DT's inference performance reaches 86% of H200

On November 18, the AI content account Max For AI stated that a document titled "Ascend 950DT Performance Testing and Business Benefit Calculation Report" was circulating in computing power-related communities. The document is dated October of this year and shows content from an internal testing report of China Mobile. According to the document, on October 10, a joint team from multiple departments of China Mobile went to Huawei's Lianqiu Lake R&D Center to conduct tests on the Zhiyuan GLM 5.2 and DeepSeek V4 Pro using a 64-card Ascend 950DT super node.The performance data provided in the document is based on the GLM 5.2, 16K input / 2K output scenario. In terms of total throughput per card, the Ascend 910B4 achieved 169 tokens/s, the 910C achieved 483 tokens/s, and the 950DT achieved 1951 tokens/s, which is approximately 11.5 times that of the 910B4 and 4 times that of the 910C. The document cites vLLM data, estimating the NVIDIA H200 in the corresponding scenario to be about 2200 tokens/s, which corresponds to the 950DT being about 89% of the H200; another set of 7K/2K data is about 82%. The final conclusion of the document states that the inference performance of the 950DT reaches 86% of the H200.In terms of cost estimation, the document lists the prices of 8-card servers based on bulk purchasing: the 950DT is approximately 2.97 million yuan, the 910C is approximately 2.15 million yuan, and the 910B4 is approximately 650,000 yuan. Including equipment, maintenance, supporting facilities, and IDC, and amortizing over five years, the monthly costs are approximately 63,600 yuan, 48,700 yuan, and 16,100 yuan, respectively. Based on this, the document calculates that the unit cost inference performance of the 950DT is about 3.1 times that of the 910C and 2.9 times that of the 910B4. The bare metal rental is estimated at 84,300 yuan per month for the 950DT and about 130,000 yuan for the H200, stating that the cost-performance ratio of the 950DT is about 30% higher than that of the H200.

NVIDIA has stopped delivering licensed Hopper products to China, with related sales of 2.4 billion dollars in the third fiscal quarter

According to Bloomberg, NVIDIA has stopped delivering Hopper data center products that received U.S. export licenses to China. Interim CEO and CFO Colette Kress stated during the earnings call that NVIDIA sold $2.4 billion worth of licensed Hopper data center products in China in the third quarter of fiscal year 2027, and has now halted related shipments. She did not specify the reason for the supply halt. NVIDIA's fourth quarter revenue guidance does not include data center computing business revenue from the Chinese market.Previously, NVIDIA had sought permission from the U.S. government to sell more advanced Blackwell chips to China, but was not approved. The Trump administration had allowed NVIDIA to export some Hopper architecture H200 chips to specific Chinese customers. The Hopper architecture to which the H200 belongs is two generations earlier than NVIDIA's current most advanced Rubin architecture. Before his departure last week, Jensen Huang had been a major proponent of NVIDIA's expansion of chip sales to China.On November 16, the U.S. Department of Commerce's Bureau of Industry and Security issued a temporary final rule to strengthen restrictions on advanced AI chips and remote access to servers, aimed at preventing companies subject to U.S. export restrictions from obtaining advanced computing power through third-country data centers. NVIDIA did not attribute this supply halt to the new regulation. On November 17, NVIDIA's board announced the establishment of a National Security and Export Control Oversight Committee, responsible for overseeing the company's national security policies, export control compliance, and global trade risk management. Chairman Harvey Jones stated that national security factors are changing the market in which NVIDIA operates, and the new CEO needs to respond cautiously.
app_icon
ChainCatcher Building the Web3 world with innovations.